Verification of Binarized Neural Networks via Inter-neuron Factoring: (Short Paper)

Abstract

Binarized Neural Networks (BNN) have recently been proposed as an energy-efficient alternative to more traditional learning networks. Here we study the problem of formally verifying BNNs by reducing it to a corresponding hardware verification problem. The main step in this reduction is based on factoring computations among neurons within a hidden layer of the BNN in order to make the BNN verification problem more scalable in practice. The main contributions of this paper include results on the NP-hardness and hardness of PTAS approximability of this essential optimization and factoring step, and we design polynomial-time search heuristics for generating approximate factoring solutions. With these techniques we are able to scale the verification problem to moderately-sized BNNs for embedded devices with thousands of neurons and inputs.